Add DPDA-GA-AASIST submission for LRLspoof (1-SRR=22.4022)

submissions/dpda-ga-aasist.yaml

@@ -0,0 +1,55 @@
+schema_version: 4
+system:
+  name: DPDA-GA-AASIST
+  slug: dpda-ga-aasist
+  description: DPDA-GA (dual-path data-augmented training with gradient alignment) on a fully fine-tuned
+    wav2vec 2.0 XLS-R 300M front-end with an AASIST spectro-temporal graph-attention back-end. Each training
+    utterance is seen as a clean view and a RawBoost-augmented view; the two views' gradients are aligned
+    with PCGrad (conflicting components projected out) before the update. Trained on ASVspoof2019 LA train
+    only; the weights are the authors' 5-checkpoint average exactly as their main.py computes it, (2*best_1
+    + best_2 + best_3 + best_4)/5 (a state_dict aliasing bug drops best_0 and counts best_1 twice; these
+    are the weights behind the paper's numbers). FP32, deterministic first-66800-sample window (shorter
+    clips repeat-padded).
+  code: https://github.com/ductuantruong/dpda_ga
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DPDA-GA-AASIST/blob/f4b395c946eee1d28be9974cfb10da5a0e08ed61/dpda_ga_aasist_avg5.pth
+  params_millions: 317.8378
+  paper:
+    arxiv_id: '2509.20682'
+    url: https://arxiv.org/abs/2509.20682
+    bibtex: "@inproceedings{truong2026addressing,\n  title     = {Addressing Gradient Misalignment in\
+      \ Data-Augmented Training for Robust Speech Deepfake Detection},\n  author    = {Truong, Duc-Tuan\
+      \ and Liu, Tianchi and Li, Junjie and Tao, Ruijie and Lee, Kong Aik and Chng, Eng Siong},\n  booktitle\
+      \ = {ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing\
+      \ (ICASSP)},\n  year      = {2026},\n  doi       = {10.1109/ICASSP55912.2026.11464739}\n}\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
+  split: test
+scores:
+  srr_complement: 22.402234679707924
+  n_trials: 1304455
+  n_skipped: 286
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: -2.575391256497175
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DPDA-GA-AASIST/resolve/9ffaa984f47218bc56e1833c2ff36d7bde799e6f/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: 543ab2a83d35ea50a1cbe5c4199e1aeeab1a1fa3a054575b2d956ccff4c7e522
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-01'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-01'
+notes: 'DPDA-GA XLS-R+AASIST from ductuantruong/dpda_ga (MIT, ICASSP 2026), PCGrad variant. Checkpoint
+  = HF tuantd/dpda_ga @ 531554e pcgrad/xlsr_aasist/best/best_{0..4}.pth averaged exactly as the source
+  main.py (--average_model True --n_average_model 5) computes it: (2*b1+b2+b3+b4)/5, because `sd = model.state_dict()`
+  aliases the params (make_avg_mainpy.py); strict load. Deterministic first-66800-sample window; shorter
+  clips repeat-tiled (source utils.pad); no resampling or normalisation. score = logit for class 1 (bona
+  fide); higher = more bona fide. Deviation: the source''s SSLModel.extract_feat re-enables train() mode
+  on XLS-R on a device/dtype mismatch; patched out so the whole net runs in eval mode. Paper (avg-5):
+  ASVspoof2021 DF 2.13% EER, In-the-Wild 5.42% EER.'